Property-based testing with Hypothesis across the numeric surface

[aarmey]

Summary

Extends property-based testing (Hypothesis) across the rest of the library's numeric surface, replacing fixed seeded-rng fixtures with generated shapes/values/densities where that's a strict generalization, and removing the tests it makes redundant.

Added tests/_hypothesis_strategies.py (shared dense_matrices, signed_dense_matrices, axis_key strategies, plus the deadline=None/too_slow-suppressed settings needed because the compress/matmul/select kernels are @numba.njit(cache=True)) and:

• test_property_ops.py — scalar mul/div, neg, transpose, log1p, matvec/matmat vs. a dense reference.
• test_property_reductions_arith.py — per-axis sum/mean/max/min and elementwise +/-/multiply.
• test_property_indexing.py — general (row_key, col_key) indexing (slices, fancy lists with duplicates/negatives, boolean masks) against dense reference indexing. Generalizes the duplicate-index "fan out" regression and the manual 25-iteration fuzz loop that used to cover it by hand.
• test_property_transpose.py — _transpose_major correctness (matches dense, nnz-preserving, involutive); replaces test_construct.py.
• test_property_normalization.py — the default normalized() view, every RECIPES entry, and select(), against a plain-numpy reference.

That last one caught a real gap in the test oracle, not the library: the reference formula didn't mirror _compute_row_scale's target <= 0.0 guard (a non-positive median row total skips depth normalization entirely). The old fixed-seed fixtures never happened to construct a matrix that hit it; Hypothesis found it within the first run. Fixed in the new reference.

Removed

test_roundtrip.py and test_construct.py (fully subsumed, deleted outright); trimmed the redundant parts of test_ops.py, test_reductions_and_arith.py, test_indexing.py, test_general_indexing_nnz_astype.py, test_select_minor_fanout.py, test_vcs_norm.py, test_vcs_norm_recipes.py, and test_norm_selection.py — keeping their error-path, structural, memory-budget, and identity/caching tests, which aren't "holds for any input" properties.

Deliberately left alone

Not good property-testing candidates: constructor validation (test_base.py), dtype-narrowing/boundary tests keyed on INT32_MAX (test_index_dtypes.py), tracemalloc memory-budget tests (test_reduction_memory.py, and the chunk-budget mechanics in test_chunked_transpose.py), and AnnData/HDF5/Zarr I/O, categorical encoding, and compression-backend dispatch (test_anndata.py, test_anndata_class.py, test_compression.py, test_metadata_encoding.py, test_rapid_load.py, test_ivcsc.py) — these test file-format/I/O contracts and specific error messages, not numeric properties.

Test plan

• uv run ruff check . / uv run ruff format --check (on touched files) / uv run ty check all clean
• uv run pytest -q — 1189 passed, 45 skipped